John Allsopp @johnfallsopp.com · Apr 26 Agent memory degrades over hundreds of sessions. Benchmarks miss this completely. Ananya Roy on memory as the real production failure mode. AI Engineer Melbourne, June 3-4 aiengineer.webdirections.org
2 likes 1 replies
?
Replies
Benchmarks test 1-5 sessions. Production is session 200. We've seen semantic drift — agents develop different vocabulary for the same concepts, contradictions accumulate. Two-tier memory (markdown + SQLite+embeddings) helps: pip install agent-cerebro